Researchers at Amazon have reportedly conducted a jailbreak that led to the Trump administration imposing export restrictions on Anthropic’s advanced artificial intelligence models, according to the Wall Street Journal. Anthropic has since blocked public access to its Fable 5 and Mythos 5 models following a directive from the Commerce Department that prohibits their use by foreign nationals and users outside the United States.
Katie Moussouris, CEO of cybersecurity firm Luta Security, stated that the jailbreak research involved using prompts to extract information about security vulnerabilities from Anthropic’s model. She received a copy of the report from Anthropic and noted that the information disclosed would be more beneficial for cybersecurity professionals than for malicious actors. Moussouris questioned the assessment made by the White House regarding the threat posed by this research, calling it an overreaction.
Amazon did not confirm its involvement in the jailbreak testing but acknowledged that it regularly consults with the government on cybersecurity matters. An Amazon spokesperson stated that it is common for governments to seek their counsel on potential security risks.
Commerce Secretary Howard Lutnick informed Anthropic CEO Dario Amodei that the Fable 5 and Mythos 5 models would be subject to export controls, which restrict access for customers outside the U.S. and foreign nationals within the country. Anthropic believes the government's decision relates to a recently identified jailbreak technique that allows users to bypass safety measures in the models, describing the vulnerabilities uncovered as minor and previously known.
Anthropic has stated that it is complying with the government's directive but disagrees with the reasoning behind the order, asserting that the finding of a narrow potential jailbreak should not necessitate recalling a commercial model used widely.